Magnitude‐scaling rates (MSRs; the rates of increase in predicted response spectra with increasing moment magnitude) are evaluated for three ground‐motion prediction equations for response spectra from subduction interface earthquakes, including two empirical models developed for data from Japan and a model based on synthetic records generated by using a stochastic finite‐fault model. MSRs vary significantly among the three models, and the difference between the two empirical models is unacceptably large. A set of 2100 strong‐motion records from subduction interface events with a magnitude of 6.5 or larger from Japan, including the 11 March 2011 magnitude 9 earthquake, were compiled. The earthquakes were grouped according to magnitude, so that the magnitude spread in each group is less than 0.2 magnitude units. Each earthquake group was treated as a single event with magnitude equal to the average magnitude for the group. A simple attenuation model was fitted to the normalized and grouped data. The model has a constant term for each earthquake group to represent the effect of magnitude. Three separate functions of magnitude (a linear model for events with a magnitude greater than 7, a bilinear model, and a curved model) were then fitted to the constants, and MSRs were derived from these functions. At short periods, the derived MSRs are only a fraction of those from two of the three attenuation models. At spectral periods greater than 0.5 s, the derived MSRs are between about 1/3 and 1/2 of those of the two empirical models but are closer to those based on a set of synthetic records.